1. The red light countdown vibration sound effect and the red light vibration sound effect should be unified to avoid confusion with the green light countdown sound effect. 2. Cancel the performance mode and only retain the energy-saving mode. The next version will cancel this button.

Red light countdown test run

Enhance floor recognition, and the same button as floor recognition. Support floors of any color. When using it, it should be one arm's distance from the floor number. The original floor recognition supports red floor numbers. The distance is unlimited.

The character recognition button is renamed the text recognition button. The scope of text recognition is expanded to the full field of view. Add the upline to pause the text and voice broadcast, and continue the text and voice broadcast down. The mobile phone turns to empty the voice announcement.

Character recognition close-up focus optimization.

Algorithm optimization, add level 2 energy-saving mode The performance mode is 50% more energy-saving than the previous version. Compared with the previous version, the energy-saving mode saves 75% energy.

Fix the problem of no-focus

The reading model is extended to 0-55. Adapt to the green and yellow countdowns. Update the floor model to 0731. Add -1, -2 floor broadcast. Update the green light model to 0731. Add pictures of traffic lights in Taiwan.

All buttons add the VoiceOver function. Floor model update, read second model update. Taiwan's second-reading model has been updated.

Floor model update Fix the default blind wand recognition range Improve the quality of bus inspection pictures

Important safety update, blind cane distance calibration. Emergency bug fix, the second open app button disappeared. Bus detection has been improved.

1. bug fix 2. Add all button setting memory functions to keep the last settings the next time they are turned on.

Floor model update Reading second model update Taiwan model update

恢复前置盲杖击球声，击球声可选。增加低频音效。 楼层模型更新，数据增加50%.

The floor identification model has been greatly updated, the training accuracy is 99%, and the practice needs to be tested. The bus route recognition function is tested.

The blind wand adds 10 optional sound effects The sound effect of the blind cane is designed as a stepless variable speed. The nearest 0.3 meters will continue to have a similar sonar radar effect. Increase the front camera floor recognition. Floor model update.

1. Floor recognition model update 2. Adjust the logic so that the zebra crossing and the reading can be broadcast at the same time. 3. Three-in-one button plus text for easy reading 4. Compass button display direction and angle. 5. The rhythm of the blind wand is adjusted to level 5: 0.15 meters-0.4 meters 7 times/second 0.4m-0.8m 6 times/second 0.8m-1.2m 4 times/second 1.2 meters-1.6 meters 3 times/second 1.6m-2.0m 2 times/second

The new function floor broadcast is in trial operation, hoping that users can use it more and accelerate the evolution. The accuracy of red light training reached four nine:99.99% for the first time. The accuracy of green light training reaches three nines: 99.97% The zebra crossing model is 100 times larger, and the zebra crossing is identified as left and right.

Rhythm adjustment of the blind cane: 7 times per second from 0.15 meters to 0.8 meters, 4 times per second from 0.8 meters to 2 meters, and silent for more than 2 meters. Reading seconds is extended from yellow numbers to green numbers. Test only.

Calibration of the length and width of the front blind stick. The running rhythm has been doubled. Increase the object recognition of chairs and sofas. The accuracy of green light training has been increased to 99.97, and the practice needs to be tested. The accuracy of red light training has been improved to 99.90, and the practice needs to be tested. The interface reduces the traffic light switch.

Front camera blind cane - deep sensing camera 30,000-dot array structure light. The algorithm optimization program has been reduced from a peak of 135 megabytes to 24 megabytes. Use of the front blind stick: After the camera switch at the bottom right switches to the front blind stick, the mobile phone screen moves forward, and the sound of hitting the ball changes according to the distance of 0.1-3m. 0.1-0.7 meters: 7 times per second 0.7-1.5 meters: 4 times per second 1.5-2.5 meters: 2 times per second 2.5-3.0 meters: 1 time per second

The red light model is reduced by 4 times, and the training accuracy is increased to 99.8%. The countdown model is reduced by 4 times, and the training accuracy is increased to 99.3%. The app is reduced from the peak 135M to 27M.

The green light model was reduced by 4 times, and the training accuracy rate was increased from 99.3% to 99.9%. Practice needs to be tested. Prepare the front camera function.

Optimize the accuracy of object recognition and usage scenarios. Fix the user feedback voice problem. Improve the visual imaging effect. Repair the reward code. Repair the evaluation code.

The green light model has been updated, and the training time has been increased by 3 times. Increase rewards for public welfare. The reward fee will all be used for the traffic light app. The priority is as follows: server rental, developer account annual fee, improvement of 5 models, data annotation, testing, and new function development.